The Investor

Two acquirers chasing the same asset tells you more than either price does, which is worth remembering before we get to the number. Stripe is in talks to buy OpenRouter — software that routes each query to the cheapest or best available AI model — for roughly $10 billion, and Databricks is circling the same company, per The Information. That is a 7.7x markup on OpenRouter's $1.3 billion last mark. When a payments giant and a data platform both bid up the same orchestration layer, the market is pricing model arbitrage as a control point, not a wrapper.

The headline number hides a tension, or rather the more interesting version of one. The exact function OpenRouter sells is being rebuilt as a free feature inside the tools developers already live in. Cursor shipped a user-facing router — cost, intelligence, or balance — that it claims cuts spend 60%, joining Factory and Ramp, per The Information AM and ben's bites. Meta is building a competing router internally and may release it publicly. So one set of buyers values routing at ten billion dollars while another set of builders drives its price toward zero. Both cannot be right about where the durable value sits.

Why the divergence is resolvable

This is probably wrong, but the reconciliation is that routing-as-a-feature and routing-as-infrastructure are different products. A router bolted into one coding tool serves that tool's users. A neutral switchboard sitting between thousands of apps and hundreds of models — with the usage data, billing relationships, and provider integrations that implies — is a tollbooth. Stripe is not paying 7.7x for the routing algorithm. It is paying for the position between demand and supply, and the payment rails that position throws off. That thesis holds only as long as enterprises want a neutral intermediary rather than their harness vendor's captive one.

Where the sources agree: three separate Information desks, AINews, and ben's bites all treat the orchestration layer as the freshest repricing event in AI infra this cycle. Where they diverge: the routing-as-commodity read (ben's bites, The Information AM) versus the routing-as-tollbooth read (The Information Dealmaker). That split is the whole trade. The comp is at its freshest the moment before the deal signs.

The adjacent signal is the IPO pipeline this repricing feeds. AlphaSense — $700M ARR growing 40%, accelerated by AI features — has engaged advisors, and behind it sit Ramp at $44B, Vercel at $9.3B (September), Verkada (30% growth, Nvidia-backed), Canva (2027), and Kalshi (2027-28). AlphaSense's eventual public multiple becomes the reference price for AI-native vertical SaaS. The OpenRouter print becomes the reference for the orchestration layer. Both will anchor private marks that have not caught up yet.

The thesis looked fundable right up until it didn't. When an OpenAI model breached Hugging Face, AI containment and red-teaming read like a clean new category, the kind investors like because it has a villain and a moat in the same sentence. Then the incumbents answered, and not by acquiring anyone — they shipped the capability for free. Anthropic launched a native Claude Security Plugin that scans codebases pre-commit from the terminal; AWS shipped GuardDuty Investigations; Cisco released Antares open-weight vulnerability models; Empirical shipped CVE prediction, most of it inside a 48-hour window, per Cyberpresso and TLDR InfoSec. Six independent startups — Perfai, Playground, Astra, OpenBox, BestDefense.io, Sequirly — now compete against features that cost nothing.

This is not a security story. It is the dominant move across the app layer:

• Code review: GitHub Copilot, Cursor, and Claude Code bundle it natively; Uber, Cloudflare, Faire, and HubSpot built in-house and reportedly did better; SonarQube absorbed Gitar. CodeRabbit, Greptile, and Qodo are the exposed standalones (The Pragmatic Engineer).
• Presentation AI: Google embedded Gemini in Slides, drafting decks from a user's own files — aimed straight at Gamma, Tome, Beautiful.ai, and Canva AI (Simplifying AI).
• Voice agents: OpenAI's Presence enters the lane held by Sierra, Cresta, Parloa, and PolyAI, so far with no disclosed pricing.
• AI-visibility (GEO/AEO): AgencyAnalytics bundled a free AI Tracker with MCP integration, undercutting the standalones (TLDR Marketing).

Where value is actually accruing

The more useful half of the trade is the mirror image. Two moats survived the bundling test. The first is proprietary data: a16z disclosed that EliseAI now runs 1-in-6 US apartments, with multi-year renter-conversation data no competitor can reproduce, and Anthropic wired Claude to its own Economic Index. The second is orchestration-plus-workflow: Wordsmith, running OpenAI, Anthropic, and Google at once, saved Belron — a $7.6B enterprise — roughly $400K in three months on legal work, cutting contract cycle time from an hour to five minutes (Applied AI). AlphaSense's $700M ARR growing 40% is the same lesson at scale, a vertical app accelerating rather than shrinking while the labs ship competitors.

The through-line is that a thin wrapper is now a feature waiting to be absorbed, while a company owning a unique dataset or the enterprise account keeps the value the labs supply as commoditized inference. The caveat worth pricing: bundling gets announced before it works — Belron has not cut headcount, and free native features often ship worse than the point solution they threaten.

Start with the number IVP is not leading with. The 2021 fund marks at 1.7x TVPI, top quartile on paper, and has returned 0.02x in realized DPI — which is to say essentially no cash back to LPs, per Newcomer's review of the confidential Fund 19 materials. In a frozen exit market that mark is a promise, not a distribution. The pitch instead anchors on a 31.1% net IRR since 1980, a figure held up almost entirely by the 1996 fund (6.7x DPI, 94.5% IRR). Any manager leaning on a 45-year inception number is pointing the eye somewhere other than the recent vintages.

The recent vintages tell a top-half story, which is a different story. The 2015 fund, IVP's best of the modern era at 2.0x DPI, posted a 22.7% IRR, trailing the top-5% benchmark of 34.3% by roughly twelve points. The 2018 and 2024 funds land top-half, not top-quartile. And yet the firm is raising $1.8 billion — back to the 2021 peak, up from $1.6B in 2024 — while asking premium economics: fees stepping from 2% to 2.25%, carry from 25% to 30% once a 2.5x hurdle clears. Premium pricing on median performance is an interesting puzzle.

The access ceiling

The harder problem is what IVP does not own. It missed both OpenAI and SpaceX, the two largest private value creators of the cycle, and its Anthropic stake is not among the largest. The AI story rests on Perplexity, Baseten, ClickHouse, Chainguard, and Abridge — a respectable roster, though not the names setting the cycle's dispersion. Which reframes brand-name growth-stage VC as capped-alpha beta, unless the manager holds meaningful stakes in the marquee winners.

Credit where it is owed: the 2021 capital discipline was genuinely shrewd. IVP distributed $2.5B while calling only $855M, front-running the correction, and that is the real edge in the deck. The mistake would be reading capital discipline as top-tier access. They are different products. Only one of them justifies top-5% pricing.

The one path where the pricing looks fair in hindsight: the 2021 paper marks harden into real DPI. That needs the exit window to open, which the OpenRouter and AlphaSense repricing elsewhere in this briefing suggests is cracking, but has not opened yet.